Meta bets big on Amazon's homegrown CPUs for AI workloads

In a move that underscores the shifting dynamics of AI infrastructure, Meta has agreed to deploy millions of AWS Graviton chips to power its growing artificial intelligence operations. Amazon announced the deal on Friday, marking a significant win for its in-house silicon strategy.

The Graviton is an ARM-based central processing unit, not a GPU. While GPUs remain essential for training large models, the rise of AI agents—which handle real-time reasoning, code generation, search, and multi-step task coordination—is driving demand for CPUs optimized for these compute-intensive workloads. Amazon says its latest Graviton was purpose-built for such AI inference tasks.

For Meta, the deal deepens its reliance on AWS after a period of diversification. Last August, Meta signed a $10 billion, six-year agreement with Google Cloud, having previously leaned heavily on AWS alongside Microsoft Azure. The timing of Amazon's announcement, coming just as Google Cloud Next wrapped up, felt like a pointed message to its cloud rival.

Amazon also produces its own AI GPU, the Trainium, used for both training and inference. But much of that capacity is already spoken for: earlier this month, Anthropic committed $100 billion over ten years to run its workloads on AWS, with Trainium at the center, while Amazon invested an additional $5 billion in the company.

With this Meta deal, Amazon can showcase a marquee customer for its Graviton CPUs, positioning them against Nvidia's new Vera CPU—also ARM-based and aimed at agentic workloads. The difference: Nvidia sells chips directly to enterprises and cloud providers, while AWS only offers access through its cloud service.

Amazon CEO Andy Jassy recently took aim at Nvidia and Intel in his shareholder letter, arguing that enterprises demand better price-performance for AI. That puts immense pressure on Amazon's chip team to deliver—a team we toured last month in an exclusive look at their lab.

Source: TechCrunch